Predicting Mental and Neurological Illnesses Based on Cerebellar Normative Features

Abstract

Background: Mental and neurological conditions have been linked to structural brain variations. However, aside from dementia, the value of brain structural characteristics derived from brain scans for prediction is relatively low. One reason for this limitation is the clinical and biological heterogeneity inherent to such conditions. Recent studies have implicated aberrations in the cerebellum, a relatively understudied brain region, in these clinical conditions.

Methods: Here, we used machine learning to test the value of individual deviations from normative cerebellar development across the lifespan (based on trained data from >27,000 participants) for prediction of autism spectrum disorder (ASD) (n = 317), bipolar disorder (n = 238), schizophrenia (SZ) (n = 195), mild cognitive impairment (n = 122), and Alzheimer's disease (n = 116); individuals without diagnoses were matched to the clinical cohorts. We applied several atlases and derived median, variance, and percentages of extreme deviations within each region of interest.

Results: The results show that lobular and voxelwise cerebellar data can be used to discriminate reference samples from individuals with ASD and SZ with moderate accuracy (the area under the receiver operating characteristic curves ranged from 0.56 to 0.65). Contributions to these predictive models originated from both anterior and posterior regions of the cerebellum.

Conclusions: Our study highlights the utility of cerebellar normative modeling in predicting ASD and SZ, aided by 4 cerebellar atlases that enhanced the interpretability of the findings.

Keywords: Cerebellum; Machine learning; Magnetic resonance imaging; Mental illnesses; Neurological diseases; Normative modeling.

Figure 1

Overview of predicting mental and neurological illnesses. (A) The study included 5 clinical datasets: Alzheimer’s disease, autism spectrum disorder, bipolar disorder, mild cognitive impairment, and schizophrenia. The input data, together with the samples, consist of both lobular and voxelwise data. (B) Individuals without a diagnosis were divided into training and test sets to evaluate the cerebellar normative models. The deviation score (z score) measures how much an individual deviates from the norm represented by the estimated population model. (C) The analysis utilized the deviation scores derived from the cerebellar lobular and voxelwise normative models. Lobular z scores consist of 28 lobular volumes. For voxelwise, deviation scores overlaid onto existing cerebellar atlases including anatomical, task-based, hierarchical, and resting-state parcellations. This process calculated median, variance, and percentage of extreme positive and negative deviation for each atlas’ regions of interest (ROIs). Logistic regression was used for each atlas to assess the predictive value of features across all ROIs.

Figure 2

Cerebellar features moderately predict autism spectrum disorder (ASD) and schizophrenia (SZ). (A) Information from the anatomical (28 regions), task-based (10 regions), hierarchical (32 regions), or resting-state (17 regions) atlases are compiled into features that were used as predictors by the logistic regression model to make predictions. The area under the receiver operating characteristic curve (AUROC) serves as an important measure in evaluating the performance of a binary classifier, representing a trade-off between the classifier’s sensitivity (true positive rate) and specificity (true negative rate). The reliability and robustness of the AUROC were assessed by computing them over 1000 permutations, which aids in determining whether the classifier’s performance is statistically significant or due to random chance. (B–D) The values that survived multiple comparison are shown. AD, Alzheimer’s disease; BD, bipolar disorder; MCI, mild cognitive impairment.

Figure 3

Different regions show distinct feature importance (FI) across atlases in autism spectrum disorder (ASD) and schizophrenia (SZ). The FI values derived from logistic regression reveal the contribution of each specific cerebellar region to predictive modeling relative to average prediction outcomes. FI values accentuate distinct cerebellar regions with unique predictive capabilities as identified in lobules, anatomical, task-based, hierarchical, and resting-state atlases through voxelwise analysis. Features that remained significant after adjustments for multiple comparisons of the area under the receiver operating characteristic curve are shown.